Three-dimensional deep learning nomogram for post-plan quality assurance in high-dose-rate prostate brachytherapy

Classical nomograms for high-dose-rate (HDR) prostate brachytherapy quality assurance (QA) evaluate agreement between planned and expected total dwell time but do not directly assess the three-dimensional treated volume. This study's aim was to develop a deep learning-based 3D nomogram to predict institution-specific clinical 100% isodose volumes from anatomy and dwell positions for post-plan QA.

A total of 2394 single-fraction 15 Gy HDR prostate brachytherapy plans were included. Input label maps encoded the 3D prostate, urethra, rectum, and dwell positions. The prediction target was the clinical 100% isodose volume. A five-fold cross-validation strategy was used for model development, with 2155 plans used for training and validation and 239 plans held out for testing. A 3D U-Net was trained using combined Dice and binary cross-entropy loss. Agreement was evaluated using prostate V100%, 100% isodose volume, Dice similarity coefficient, and average symmetric surface distance (ASSD). Validation-derived thresholds were used for outlier screening and targeted review.

The model predicted prostate V100% with a mean error of 0.73 percentage points. Predicted and clinical 100% isodose volumes showed strong agreement, with R² = 0.990, slope = 0.99, and intercept = 0.77 cm³. Median Dice similarity coefficient was 0.960, and median ASSD was 0.48 mm. A 95% threshold flagged 6.3% of plans as atypical relative to validation data.

A deep learning-based nomogram predicted an institution-specific 100% isodose volume with low coverage error and high spatial agreement. This geometry aware post-plan QA tool can complement equation-based nomograms by directly evaluating the spatial treated volume.